Xiaomi

Type: Consumer-Electronics Company / Robotics Foundation-Model Entrant

Xiaomi — best known as a consumer-electronics and smartphone maker — entered the VLA foundation-model race directly with Xiaomi-Robotics-1 (Jul 2026), a 33-author corporate-research release from the Xiaomi Robotics Team rather than an academic-lab paper. That in itself is a signal: the humanoid/VLA foundation-model competition is no longer confined to dedicated robotics labs (Physical Intelligence), GPU/simulation platforms (NVIDIA), or humanoid-hardware makers (Tesla, Figure) — a company whose core business is consumer devices is now a credible entrant on the strength of data-scale alone.

Strategic Position

Xiaomi's bet is data-scale via cheap capture infrastructure: Xiaomi-Robotics-1 pre-trains on 100k+ hours of real-world manipulation trajectories collected via UMI (Universal Manipulation Interface) devices — a portable, robot-free demonstration-capture method — combined with a scalable auto-labeling pipeline that turns raw trajectory clips into richly language-conditioned training data without manual annotation. This sits in the same "data-scale + ecosystem" family as NVIDIA's GR00T (20K hours of EgoScale video), but pushes data volume roughly 5x further and reports the scaling has not saturated — continued data/model-size increases keep improving downstream performance. See Foundation Models for Robotics for the full recipe-race comparison.

Key Contributions

• Xiaomi-Robotics-1: VLA pre-trained on 100k+ hours of UMI-collected real-world trajectories with auto-labeled language conditioning; new SOTA on RoboCasa365 (57.4% vs. 46.6% prior) and RoboDojo (20.07 vs. 13.07 prior); scaling reported as unsaturated across both data volume and model size during pre-training, with gains transferring into post-training. Evidence: moderate (single preprint, author-benchmarked; code/checkpoints not yet public) (Xiaomi-Robotics-1)

Mentioned In

• Foundation Models for Robotics — data-scale entrant in the humanoid/VLA foundation-model recipe race

Related Entities

• NVIDIA — competing data-scale + ecosystem play (GR00T)
• Physical Intelligence — competing heterogeneous-co-training play (π₀.₅)

Open Questions

• Does Xiaomi have (or plan) a hardware platform to deploy Xiaomi-Robotics-1 on, or is this a foundation-model-only research contribution aimed at licensing?
• Will Xiaomi release the promised code and model checkpoints, enabling independent verification of the reported SOTA numbers?

Changelog

• 2026-07-23 — Initial compilation from xiaomi-robotics-1-vla-scaling.